There are two classes of Decision Tree algorithm in scikit-learn library:
Below I show 4 ways to visualize Decision Tree in python:
- print text representation of the tree with
- plot with
sklearn.tree.plot_treemethod (matplotlib needed)
- plot with
sklearn.tree.export_graphvizmethod (graphviz needed)
- plot with
dtreevizpackage (dtreeviz and graphviz needed)
I will show how to visualize trees on classification and regression tasks.
Train Decision Tree on Classification Task
I will train a
iris dataset. I will use default hyper-parameters for the classifier.
from matplotlib import pyplot as plt from sklearn import datasets from sklearn.tree import DecisionTreeClassifier from sklearn import tree
# Prepare the data data iris = datasets.load_iris() X = iris.data y = iris.target
# Fit the classifier with default hyper-parameters clf = DecisionTreeClassifier(random_state=1234) model = clf.fit(X, y)
Print Text Representation
Exporting Decision Tree to the text representation can be useful when working on applications whitout user interface or when we want to log information about the model into the text file. You can check details about
export_text in the sklearn docs.
text_representation = tree.export_text(clf) print(text_representation)
|--- feature_2 <= 2.45 | |--- class: 0 |--- feature_2 > 2.45 | |--- feature_3 <= 1.75 | | |--- feature_2 <= 4.95 | | | |--- feature_3 <= 1.65 | | | | |--- class: 1 | | | |--- feature_3 > 1.65 | | | | |--- class: 2 | | |--- feature_2 > 4.95 | | | |--- feature_3 <= 1.55 | | | | |--- class: 2 | | | |--- feature_3 > 1.55 | | | | |--- feature_0 <= 6.95 | | | | | |--- class: 1 | | | | |--- feature_0 > 6.95 | | | | | |--- class: 2 | |--- feature_3 > 1.75 | | |--- feature_2 <= 4.85 | | | |--- feature_1 <= 3.10 | | | | |--- class: 2 | | | |--- feature_1 > 3.10 | | | | |--- class: 1 | | |--- feature_2 > 4.85 | | | |--- class: 2
If you want to save it to the file, it can be done with following code:
with open("decistion_tree.log", "w") as fout: fout.write(text_representation)
Plot Tree with
plot_tree method was added to sklearn in version
0.21. It requires
matplotlib to be installed. It allows us to easily produce figure of the tree (without intermediate exporting to graphviz) The more information about
plot_tree arguments are in the docs.
fig = plt.figure(figsize=(25,20)) _ = tree.plot_tree(clf, feature_names=iris.feature_names, class_names=iris.target_names, filled=True)
plot_tree returns annotations for the plot, to not show them in the notebook I assigned returned value to
To save the figure to the
Please notice that I’m using
filled=True in the
plot_tree. When this parameter is set to
True the method uses color to indicate the majority of the class. (It will be nice if there will be some legend with class and color matching.)
Visualize Decision Tree with graphviz
Please make sure that you have
graphviz installed (
pip install graphviz). To plot the tree first we need to export it to DOT format with
export_graphviz method (link to docs).
Then we can plot it in the notebook or save to the file.
import graphviz # DOT data dot_data = tree.export_graphviz(clf, out_file=None, feature_names=iris.feature_names, class_names=iris.target_names, filled=True) # Draw graph graph = graphviz.Source(dot_data, format="png") graph
Plot Decision Tree with
dtreeviz package is available in github. It can be installed with
pip install dtreeviz. It requires
graphviz to be installed (but you dont need to manually convert between DOT files and images). To plot the tree just run:
from dtreeviz.trees import dtreeviz # remember to load the package viz = dtreeviz(clf, X, y, target_name="target", feature_names=iris.feature_names, class_names=list(iris.target_names)) viz
Save visualization to the file:
Visualizing the Decision Tree in Regression Task
Below, I present all 4 methods for
DecisionTreeRegressor from scikit-learn package (in python of course).
from sklearn import datasets from sklearn.tree import DecisionTreeRegressor from sklearn import tree
# Prepare the data data boston = datasets.load_boston() X = boston.data y = boston.target
To keep the size of the tree small, I set
max_depth = 3.
# Fit the regressor, set max_depth = 3 regr = DecisionTreeRegressor(max_depth=3, random_state=1234) model = regr.fit(X, y)
text_representation = tree.export_text(regr) print(text_representation)
|--- feature_5 <= 6.94 | |--- feature_12 <= 14.40 | | |--- feature_7 <= 1.38 | | | |--- value: [45.58] | | |--- feature_7 > 1.38 | | | |--- value: [22.91] | |--- feature_12 > 14.40 | | |--- feature_0 <= 6.99 | | | |--- value: [17.14] | | |--- feature_0 > 6.99 | | | |--- value: [11.98] |--- feature_5 > 6.94 | |--- feature_5 <= 7.44 | | |--- feature_4 <= 0.66 | | | |--- value: [33.35] | | |--- feature_4 > 0.66 | | | |--- value: [14.40] | |--- feature_5 > 7.44 | | |--- feature_10 <= 19.65 | | | |--- value: [45.90] | | |--- feature_10 > 19.65 | | | |--- value: [21.90]
fig = plt.figure(figsize=(25,20)) _ = tree.plot_tree(regr, feature_names=boston.feature_names, filled=True)
Please notice, that the color of the leaf is coresponding to the predicted value.
dot_data = tree.export_graphviz(regr, out_file=None, feature_names=boston.feature_names, filled=True) graphviz.Source(dot_data, format="png")
from dtreeviz.trees import dtreeviz # remember to load the package viz = dtreeviz(regr, X, y, target_name="target", feature_names=boston.feature_names) viz
From above methods my favourite is visualizing with
dtreeviz package. I like it becuause:
- it shows the distribution of decision feature in the each node (nice!)
- it shows the class-color matching legend
- it shows the distribution of the class in the leaf in case of classification tasks, and mean of the leaf’s reponse in the case of regression tasks
It would be great to have
dtreeviz visualization in the interactive mode, so the user can dynamically change the depth of the tree. I’m using
dtreeviz package in my Automated Machine Learning (autoML) Python package
mljar-supervised. You can check the details of the implementation in the github repository. One important thing is, that in my AutoML package I’m not using decision trees with
max_depth greater than
4. I add this limit to not have too large trees, which in my opinion loose the ability of clear understanding what’s going on in the model. Below is the example of the markdown report for Decision Tree generated by